Mask3D: Mask Transformer for 3D Semantic Instance Segmentation

Jonas Schult; Francis Engelmann; Alexander Hermans; Or Litany; Siyu Tang; Bastian Leibe

Source-linked AI summary

Mask3D: Mask Transformer for 3D Semantic Instance Segmentation

Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, Bastian Leibe

arXiv:2210.03105v2Oct 6, 2022cs.CV

TL;DR

3D instance segmentation methods commonly depend on voting and geometric grouping, while Transformer-based methods had not addressed this task. Mask3D uses multi-scale Transformer-refined instance queries to directly predict semantic masks, achieving state-of-the-art results across several datasets. Its reported limitation is that similar, distant objects can sometimes be merged.

Problem
Existing 3D instance-segmentation approaches rely on hand-selected voting properties, grouping heuristics, and proxy losses instead of directly predicting instance masks.
Method
Mask3D uses Transformer decoders to iteratively refine instance queries from multi-scale point features, directly predicting semantic classes and instance masks.
Results
Mask3D achieves state-of-the-art performance on ScanNet, ScanNet200, S3DIS, and STPLS3D, including at least 6.2 mAP gains on ScanNet and 11.2 mAP on STPLS3D.
Takeaways & Limitations
Mask3D provides a Transformer-based alternative that avoids center voting, nonmaximum suppression, and grouping heuristics for 3D semantic instance segmentation.
Takeaways & Limitations
Similar distant objects can be merged when they produce similar learned point features, a failure less likely for methods encoding geometric priors.

Abstract

from arXiv · show

Modern 3D semantic instance segmentation approaches predominantly rely on specialized voting mechanisms followed by carefully designed geometric clustering techniques. Building on the successes of recent Transformer-based methods for object detection and image segmentation, we propose the first Transformer-based approach for 3D semantic instance segmentation. We show that we can leverage generic Transformer building blocks to directly predict instance masks from 3D point clouds. In our model called Mask3D each object instance is represented as an instance query. Using Transformer decoders, the instance queries are learned by iteratively attending to point cloud features at multiple scales. Combined with point features, the instance queries directly yield all instance masks in parallel. Mask3D has several advantages over current state-of-the-art approaches, since it neither relies on (1) voting schemes which require hand-selected geometric properties (such as centers) nor (2) geometric grouping mechanisms requiring manually-tuned hyper-parameters (e.g. radii) and (3) enables a loss that directly optimizes instance masks. Mask3D sets a new state-of-the-art on ScanNet test (+6.2 mAP), S3DIS 6-fold (+10.1 mAP), STPLS3D (+11.2 mAP) and ScanNet200 test (+12.4 mAP).

I. INTRODUCTION

Mask3D addresses 3D semantic instance segmentation by directly predicting labeled instance masks with a Transformer, avoiding voting and hand-crafted grouping components used by prior approaches.

3D semantic instance segmentation outputs binary foreground masks over input points together with semantic object labels.
Current approaches often rely on voting for hand-selected geometric properties and heuristics for clustering votes, rather than directly predicting instance masks.
Transformer-based methods had not previously been applied to 3D instance segmentation, and existing 3D Transformer approaches lagged current state-of-the-art performance.
Mask3D represents each object instance with a query that encodes semantic and geometric information and is decoded into a class and instance feature.
Mask3D avoids center voting, nonmaximum suppression, and grouping heuristics while achieving state-of-the-art performance across ScanNet, ScanNet200, S3DIS, and STPLS3D.

II. RELATED WORK

Related 3D instance-segmentation methods include bottom-up, top-down, and voting-based approaches, with some using multi-scale features followed by offline clustering.

3D instance segmentation methods include bottom-up, top-down, and voting-based approaches.
MASC uses a multi-scale hierarchical feature backbone, but computes pairwise affinities and applies offline clustering.

III. METHOD

Mask3D combines multi-scale point features with Transformer-decoded instance queries, iteratively refining queries and directly producing semantic classes and binary instance masks.

The model combines a feature backbone with a Transformer decoder built from mask modules and query-refinement layers.
The feature backbone outputs multi-scale point features, while the decoder iteratively refines instance queries through cross-attention to those features.
A mask module combines refined queries with point features to predict a semantic class and an instance mask for each query.
Similarity scores from point-query dot products pass through a sigmoid and thresholding to produce binary masks.
Auxiliary binary masks guide masked cross-attention in the next refinement step, while semantic labels are predicted directly rather than by majority voting or grouping.
The decoder attends to the four coarsest backbone levels from coarse to fine, repeating the process three times for 12 query-refinement steps.

A. Training and Implementation Details

Mask3D trains unordered predicted instances against ground-truth instances using bipartite matching, then directly optimizes masks and classifications across decoder layers. Its confidence combines class confidence with the mean confidence of voxels inside the binarized mask.

Training objective: Bipartite graph matching establishes correspondences between unordered predicted and ground-truth instance sets during training.The matching cost combines Dice, binary cross-entropy mask, and classification losses.
Training objective: The Hungarian method efficiently finds the optimal assignment under the matching cost.The weights are λdice = λcl = 2.0 and λBCE = 5.0.
Training objective: After matching, each predicted mask is directly optimized with binary cross-entropy and Dice losses, while classification uses multi-class cross-entropy.Unassigned masks are supervised toward the no-object class with λno-obj. = 0.1.
Training objective: The overall loss supervises auxiliary instance predictions after each of the L decoder layers.
Prediction confidence: Prediction confidence multiplies semantic class confidence by a mask confidence based on the mean confidence of voxels inside the binarized mask.High-confidence instances require both confident classification and predominantly high-confidence voxels.
Query types: Non-parametric queries improve performance and allow changing the number of inference queries without retraining, trading inference speed against performance.They are initialized from sampled point-cloud coordinates using farthest point sampling.

IV. EXPERIMENTS

The experiments compare Mask3D with prior state-of-the-art methods on four public 3D instance-segmentation datasets, followed by analyses of query design and refinement settings. Qualitative results and limitations are discussed separately.

IV. EXPERIMENTS: Mask3D is compared with prior state-of-the-art methods on four publicly available 3D indoor and outdoor datasets.
IV. EXPERIMENTS: The ScanNet comparison reports mean average precision across IoU thresholds over 18 classes on validation and test sets.Inference speed is averaged over validation scenes on a TITAN X GPU, excluding postprocessing.
IV. EXPERIMENTS: The study analyzes query types, the number of query-refinement steps, and the number of inference queries before presenting qualitative results and limitations.

A. Comparing with State-of-the-Art Methods

Mask3D is evaluated across ScanNet, ScanNet200, S3DIS, and STPLS3D using dataset-specific instance-segmentation metrics. It outperforms prior work by substantial mAP margins across these benchmarks and retains strong performance under class imbalance without architectural or training modifications.

Datasets and metrics: ScanNet contains 3D reconstructed indoor scenes with semantic and instance labels for 18 object categories, evaluated using mean average precision.Its splits contain 1202 training, 312 validation, and 100 hidden test scenes.
Datasets and metrics: S3DIS contains 272 scans across six areas and 13 classes, evaluated with mAP and mean precision/recall at 50% IoU.Results include Area-5 and 6-fold cross-validation settings.
Datasets and metrics: ScanNet200 extends ScanNet with an order of magnitude more classes, testing performance under natural class imbalance and challenging long-tail categories.
Results: Mask3D outperforms prior work by at least 6.2 mAP on ScanNet, 6.2 mAP on S3DIS, 10.8 mAP on ScanNet200, and 11.2 mAP on STPLS3D.These comparisons use the most challenging reported mAP metric.
Results: Mask3D performs strongly on indoor and outdoor datasets and under challenging class imbalance without inherent architectural or training-regime modifications.ScanNet pre-training improves S3DIS Area-5 performance by 1.2 mAP.

B. Analysis Experiments

The ablations examine query initialization, query counts, decoder layers, and mask losses, showing trade-offs between performance, runtime, memory, and mask-optimization choices.

Query Types: Non-parametric queries using sampled positions outperform learned parametric queries, while adding features at sampled positions degrades performance.Queries are initialized from furthest point sampling (FPS); position-only non-parametric queries perform best among the compared variants.
Number of Queries and Decoders: Increasing trained queries from 100 to 200 slightly improves performance but increases memory use.The analysis evaluates models trained with K = 100 and K = 200 non-parametric queries sampled with FPS.
Number of Queries and Decoders: Using fewer inference queries reduces performance but speeds runtime, whereas using more produces typically less than 1% improvement.The model is evaluated with query counts differing from the training configuration.
Mask Losses: A weighted combination of binary cross-entropy and Dice losses gives the best instance-mask prediction performance.Dice loss addresses the imbalance between foreground mask points and background points, while the combination improves over binary cross-entropy alone.

C. Qualitative Results and Limitations

Mask3D produces robust qualitative ScanNet results across cluttered and challenging scenes, but can merge distant instances with similar semantics and geometry.

Qualitative Results: Mask3D shows robust results on diverse ScanNet scenes containing clutter, scanning artifacts, and numerous similar objects.Representative examples include predicted instance masks and semantic labels.
Limitations: A systematic failure merges far-apart objects when similar learned point features cause the attention mechanism to combine them into one instance.Methods that explicitly encode geometric priors are less likely to exhibit this error.

V. CONCLUSION

The conclusion presents Mask3D as a Transformer-decoder approach that learns instance queries and directly predicts semantic instance masks without hand-selected voting or grouping mechanisms.

V. CONCLUSION: Mask3D learns instance queries that combine with learned point features to directly predict semantic instance masks.The approach is introduced for 3D semantic instance segmentation.
V. CONCLUSION: The method avoids hand-selected voting schemes and hand-crafted grouping mechanisms used by voting-based approaches.The authors characterize Mask3D as an alternative to current voting-based methods.

Supplementary Material

The supplementary material describes the full Mask3D architecture and notes experiments with convolutional and Transformer-based feature backbones.

Supplementary Material: The full model uses five hierarchical feature scales and Transformer decoder layers that iteratively refine instance queries.The supplementary illustration expands the simplified main-paper model with additional feature levels and query-refinement layers.
Supplementary Material: The supplementary experiments compare convolutional and Transformer-based feature backbones.These backbone variants are referenced in relation to the full model architecture.

I. IMPLEMENTATION DETAILS

Mask3D is trained and evaluated with dataset-specific scene handling, a multiscale Transformer architecture, and optional DBSCAN postprocessing for merged instances. Additional experiments indicate its segmentation performance is not primarily due to model size.

S3DIS Specific Details: S3DIS training uses randomly cropped 6m×6m blocks, while testing disables cropping for full-scene inference.Because scenes can exceed several million points, training runs for 1000 epochs.
STPLS3D Specific Details: STPLS3D inference uses 54m×54m blocks but retains predictions relevant to the evaluated 50m×50m regions.This handling improves results by roughly 1.2 mAP.
Model Details: The model uses five backbone feature scales projected to 128 dimensions and a modified Mask2Former Transformer decoder with eight-headed attention.The backbone is Minkowski Res16UNet34C, and the decoder uses 1024-dimensional feedforward features.
Density-Based Clustering: DBSCAN is an optional postprocessing routine for splitting similarly looking instances that Mask3D incorrectly merges.Reported best scores occur near ϵ=0.9 on ScanNet and ϵ=0.6 on S3DIS-A5.
Model Size: 40.9 vs. 40.0 mAP on ScanNet validation 5 cm shows comparable performance with the smaller Res16UNet18B backbone.The experiment tests whether Mask3D’s gains originate from using more model parameters.